TGO-005

Novissi – Machine Learning on Mobile Phone Metadata for Individual Poverty Prediction and Beneficiary Targeting

Download PDF
Togo Sub-Saharan Africa Low income Operational Deployment (Limited Rollout) Confirmed

Government of Togo – Ministry of Digital Economy and Digital Transformation (MENTD)

At a Glance

What it does Prediction (including forecasting) — Vulnerability, needs and risk assessment, including predictive analytics
Who runs it Government of Togo – Ministry of Digital Economy and Digital Transformation (MENTD)
Programme Novissi Emergency Cash Transfer Programme – Phase 2 (Rural Expansion with ML Targeting)
Confidence Confirmed
Deployment Status Operational Deployment (Limited Rollout)
Key Risks Data-related risks
Key Outcomes The strongest documented case-specific outcome is that the phone-based targeting approach helped identify roughly 57,000 beneficiaries in the rural expansion phase without requiring fresh household surveys, and academic evaluation found lower exclusion errors than feasible geographic alternatives considered in context.
Source Quality 5 sources — Working paper / technical note, Academic journal article, Dataset / database, +1 more

The Government of Togo used a phone-based poverty-prediction model during the rural expansion of Novissi, its COVID-19 emergency cash-transfer programme, to help prioritise potential beneficiaries within already selected cantons. The retained evidence strongly supports that this component used anonymised mobile-phone metadata and survey-based training data to estimate poverty at individual level and to rank potential recipients for programme outreach. As with the area-targeting component, the case is best framed as a government use of a research-developed targeting tool inside a broader emergency programme rather than as a fully disclosed state-built AI product.

Novissi's second phase sought to extend support into poorer rural areas during a period when household surveys and in-person registration were difficult. The strongest source base shows that the individual-targeting model relied on approximately 150 features derived from call detail records and related mobile-phone usage patterns, trained against survey-based consumption data. With financial support from the World Bank, the research team conducted a large phone survey of roughly 10,000 individuals in September 2020, immediately prior to the rural expansion, to provide ground-truth information on living conditions. The team went to considerable lengths to ensure representativeness through adaptive survey design and the use of survey sample weights, so that difficult-to-reach populations such as the extreme poor and those living in remote villages were represented in the training data. Each survey respondent provided informed consent to participate. The mobile phone metadata used as model inputs included information about the date, time, duration, and cell tower used for calls and texts, as well as data on mobile data usage volume and mobile money transaction patterns. From these raw data the research team derived aggregate statistics of each subscriber's phone-use patterns, including features correlated with wealth such as the total volume of international phone calls and average mobile money balance. The algorithms were then used to generate a consumption estimate for each of the 5.7 million mobile subscribers in the country. Public documentation is relatively strong on the research methodology and fairness trade-offs, but weaker on the internal government operating rules used when model outputs were translated into programme decisions.

This component sat downstream of a geographic pre-selection step that identified priority cantons. Within the 100 poorest cantons, where approximately 580,000 citizens lived, the government in collaboration with GiveDirectly had secured sufficient funding to provide benefits to roughly 57,000 individuals. The model generated ranked estimates and those estimated to consume less than 1.25 US dollars per day were prioritised for Novissi transfers. The government retained control over the programme parameters, and the sources describe the model as part of a broader targeting pipeline rather than as an autonomous benefit-decision engine. Preliminary evaluation results reported by IPA indicated that assuming the goal is to reach the poorest 57,000 people in the 100 poorest cantons, the satellite-plus-phone approach was significantly more accurate than the alternative approaches available to the government, providing benefits to nearly 2.5 times as many of the poorest citizens as an occupation-based targeting approach. The CEGA case study reports a 42 percent improvement in targeting precision relative to naive geographic targeting and states that 154,238 citizens received unconditional cash transfers between December 2020 and April 2021 through the scaled approach.

Data privacy was a central concern in programme design. Neither GiveDirectly nor the Government of Togo had access to any data collected by the mobile phone operators, and neither received access to the poverty scores derived from the mobile data. Instead, the research team produced a list of eligible beneficiaries based on the poverty scores, and the government received only that list. The researchers implemented strict anonymisation, encryption, and access protocols, and UC Berkeley's Committee for the Protection of Human Subjects reviewed all research procedures. The research team also designed algorithmic audits to examine whether specific vulnerable subgroups were more likely to be excluded.

The case is important because it shows a real, high-consequence use of model-based targeting in social protection under crisis conditions, with unusually strong academic validation compared with many other public-sector examples. A public replication repository was released on GitHub with notebook-level replication for survey processing, satellite poverty mapping, machine-learning modelling, targeting simulations, and fairness analysis, with synthetic data released for the non-public CDR inputs. But the case also remains partner-heavy: much of the most detailed evidence comes from academic and development-partner materials, not from direct state-authored operational disclosure. The safest production framing is therefore that Togo operationally used a phone-metadata-based poverty-prediction model in Novissi, while leaving broader claims about comparative performance and programme-wide effects carefully bounded.

Classifications follow the DCI AI Hub Taxonomy. Hover over field labels for definitions.

Social Protection Functions

Implementation/delivery chain
Assessment of needs/conditions + enrolment primary
SP Pillar (Primary) The social protection branch: social assistance, social insurance, or labour market programmes. Social assistance
Programme Name Novissi Emergency Cash Transfer Programme – Phase 2 (Rural Expansion with ML Targeting)
Programme Type The type of social protection programme, classified under social assistance, social insurance, or labour market programmes. View in glossary Emergency Cash Transfers
System Level Where in the social protection system the AI is applied: policy level, programme design, or implementation/delivery chain. View in glossary Implementation/delivery chain
Programme Description The Novissi programme is Togo's emergency cash transfer platform launched in April 2020 in response to the COVID-19 pandemic. This case covers the individual-level phone-metadata targeting component used in the rural expansion phase, not the full Novissi delivery chain.
Implementation Type How the AI output is produced: Classical ML, Deep learning, Foundation model, or Hybrid. Affects validation, compute requirements, and governance profile. View in glossary Classical ML
Lifecycle Stage Current stage in the AI lifecycle, from problem identification through to monitoring, maintenance and decommissioning. View in glossary Monitoring, Maintenance and Decommissioning
Model Provenance Origin of the AI model: developed in-house, adapted from open-source, commercial/proprietary, or accessed via third-party API. View in glossary Adapted from open-source
Compute Environment Where the AI system runs: on-premise, government cloud, commercial cloud, or edge/device. View in glossary Not documented
Sovereignty Quadrant Classification of data and compute sovereignty: I (Sovereign), II (Federated/Hybrid), III (Cloud with safeguards), or IV (Shared Innovation Zone). View in glossary IV — Shared Innovation Zone
Data Residency Where the data used by the AI system is stored: domestic, regional, or international. View in glossary International
Data Residency Detail Additional detail on the specific data hosting arrangements and jurisdictions. CDR data held by Togolese mobile network operators; analytical processing conducted by US-based research institutions (UC Berkeley, Northwestern University) under data sharing agreements.
Cross-Border Transfer Whether data crosses national borders, and if so, whether documented safeguards are in place. View in glossary Without documented safeguards
Decision Criticality The rights impact of the decision the AI supports. High criticality requires HITL oversight; moderate requires HOTL; low may operate HOOTL. View in glossary High
Human Oversight Type Level of human involvement: Human-in-the-Loop (active review), Human-on-the-Loop (monitoring), or Human-out-of-the-Loop (periodic audit). View in glossary HITL
Development Process Whether the AI system was developed fully in-house, through a mix of in-house and third-party, or fully by an external provider. View in glossary Mix of in-house and third-party
Highest Risk Category The most significant structural risk source identified: data, model, operational, governance, or market/sovereignty risks. View in glossary Data-related risks
Risk Assessment Status Whether a formal risk assessment, informal assessment, or independent audit has been conducted for this system. Formal assessment

Risk Dimensions

Governance and institutional oversight risks
Market, sovereignty and industry structure risks

Impact Dimensions

Autonomy, human dignity and due process
Systemic and societal
  • Bias audit
  • Data minimisation controls
  • Human oversight protocol
  • Independent evaluation
CategorySensitivityCross-System LinkageAvailabilityKey Constraints
Financial and payments data: beneficiary financial behaviourSensitiveLinks data across multiple systemsCurrently available and usedMobile money transaction volumes included as features in CDR-derived dataset; GiveDirectly mobile money payment infrastructure used for disbursement
Survey and census dataPersonalLinks data across multiple systemsCurrently available and usedGround-truth consumption data from phone surveys of ~10,000 individuals used for model training and validation; three rounds of high-frequency mobile phone surveys conducted
Telecommunications and mobile dataSensitiveLinks data across multiple systemsCurrently available and usedAnonymised CDR data from Togocel and Moov Africa under data access agreements; ~150 features derived from call/SMS frequency, duration, timing, cell tower mobility, mobile money transactions, international calling, data usage; covers ~5.7 million subscribers (~70% of population)

Center for Effective Global Action (2020) 'Using AI and Digital Data to Target Cash Transfers in Togo'. Available at: https://cega.berkeley.edu/collection/ai-assisted-cash-transfers-togo/ (Accessed: 27 March 2026).

View source Working paper / technical note

Aiken, E., Bellue, S., Karlan, D., Udry, C. and Blumenstock, J. (2022) 'Machine learning and phone data can improve targeting of humanitarian aid', Nature, 603, pp. 864–870. doi:10.1038/s41586-022-04484-9.

View source Academic journal article

Laiken, E. (2022) 'togo-targeting-replication', GitHub repository. Available at: https://github.com/emilylaiken/togo-targeting-replication (Accessed: 27 March 2026).

View source Dataset / database

Innovations for Poverty Action (2021) 'Using Mobile Phone and Satellite Data to Target Emergency Cash Transfers in Togo', 12 January. Available at: https://poverty-action.org/using-mobile-phone-and-satellite-data-target-emergency-cash-transfers-togo (Accessed: 27 March 2026).

View source Working paper / technical note

World Bank (2021) 'Prioritizing the poorest and most vulnerable in West Africa: Togo's Novissi platform for social protection uses machine learning, geospatial analytics, and mobile phone metadata for the pandemic response', Results Briefs, 13 April. Washington, DC: World Bank.

View source Report (multilateral / development partner)
Deployment Status How far the system has progressed into real-world operational use, from concept/exploration through to scaled and institutionalised. View in glossary Operational Deployment (Limited Rollout)
Year Initiated The year the AI system was first initiated or development began. 2020
Scale / Coverage The scale and geographic or population coverage of the deployment. 5.7 million mobile subscribers scored; ~57,000 beneficiaries identified via ML in rural expansion phase (Phase 2); ~140,000 Phase 2 beneficiaries total; 920,000+ beneficiaries across all Novissi phases
Funding Source The source(s) of funding for the AI system development and deployment. World Bank IDA financing under the West Africa Unique Identification for Regional Integration and Inclusion (WURI) Program; IDA provided $72 million for social protection delivery systems.
Technical Partners External technology vendors, academic partners, or development partners involved. Academic research consortium (UC Berkeley, Northwestern University, IPA) developed methodology; no commercial vendor for core ML model; mobile network operators (Togocel, Moov Africa) provided CDR infrastructure; GiveDirectly executed payments.
Outcomes / Results The strongest documented case-specific outcome is that the phone-based targeting approach helped identify roughly 57,000 beneficiaries in the rural expansion phase without requiring fresh household surveys, and academic evaluation found lower exclusion errors than feasible geographic alternatives considered in context. The sources also indicate that the selected beneficiaries were poorer on average than the broader population. Broader Novissi spending, coverage, and mobile-money-account figures provide programme context, but they should not be treated as attributable solely to this individual-targeting model.
Challenges Exclusion of non-phone owners from algorithmic targeting (phone ownership is a prerequisite); potential bias against populations with atypical phone usage patterns; no comprehensive social registry exists in Togo for comparison; limited transparency on individual-level eligibility decisions.

How to Cite

DCI AI Hub (2026). 'Novissi – Machine Learning on Mobile Phone Metadata for Individual Poverty Prediction and Beneficiary Targeting', AI Hub AI Tracker, case TGO-005. Digital Convergence Initiative. Available at: https://socialprotectionai.org/use-case/TGO-005 [Accessed: 1 April 2026].

Change History

Created 30 Mar 2026, 08:41
by v2-import (import)